Internet Technology Letters最新文献

Implementing efficient and high-quality silkworm cocoon sorting is of great significance in the IIoT context. This study proposes an enhanced YOLOv5 deep learning model combined with data augmentation techniques to classify cocoon images. Specifically, the model achieved an accuracy (Acc) of over 97% in recognizing off-cluster cocoons, while the accuracy for on-cluster fresh cocoons decreased by more than 10%, and that for thin-shelled cocoons dropped to approximately 70%. However, after dataset enhancement, the overall recognition rate of defective and damaged fresh silkworm cocoons reached 85.2% and the thin-shelled cocoons increased by 7%, which meets the requirements of real-time intelligent identification and classification. This research demonstrates the potential of the IIoT in improving efficiency and quality control in the silk industry. Meanwhile, the observed performance fluctuations indicate that further model optimization is needed to improve the robustness of the model under on-cluster conditions.

The construction industry is increasingly adopting technological innovations such as building information modeling (BIM), 5G, and computer vision to transform traditional construction management practices. These advancements represent a prominent research trend and are attracting significant attention in practical applications. However, current construction site safety monitoring systems still face several challenges, including delayed BIM model updates, insufficient multi-source data integration, and limitations in edge computing resources due to on-site constraints, all of which impede the stable performance of safety monitoring. To address these issues, this paper proposes a multimodal real-time anomaly monitoring model named MM-BIM, based on 5G edge intelligence and lightweight BIM. By leveraging 5G for transmitting laser point clouds and IoT data, the system enables incremental BIM updates at the edge, with model lightweighting achieved through Draco data compression. Integrating YOLOv8, transformer, and sensor knowledge graphs allows for millisecond-level detection of risks and structural deviations. Experimental results demonstrate a 94.2% mAP in safety behavior recognition, a warning response latency of 218 ms, reliable monitoring performance under high-wind conditions, and a low false alarm rate.

With the rise of the animation industry, animation courses in universities are becoming increasingly popular. However, traditional teaching focuses too much on software operation and neglects animation principles (such as the 3D creation process), resulting in insufficient practical ability of students and an urgent need for reform. To cultivate students' comprehensive practical ability in animation design, this article constructs a teaching platform based on artificial intelligence technology, integrates multiple environmental resources, and promotes the transformation of the teaching mode toward ability cultivation. In addition, the platform adopts a modular design to construct a corpus module with a high-frequency word filtering effect, and proposes a character animation generation module based on the ontology model library to automatically generate scripts. Experimental verification shows that the PSNR value has increased by about 12 dB, the SSIM average is 0.92, the user evaluation average score is 4.6 points (visual quality), and the PCK keypoint accuracy has an average of 93.9% (such as wrist 97.7%), which is significantly better than the comparison method. The platform constructed effectively enhances students' practical abilities, supports teaching innovation, and will expand big data architecture to optimize applications in the future.

With the rise of social media chat platforms, spam, particularly in the form of text-embedded images has become increasingly disruptive. This study proposes a lightweight, reference-based, unsupervised spam detection method that offers both computational efficiency and high adaptability. Unlike traditional supervised models, the approach leverages string similarity algorithms to detect phrases resembling user-defined spam references. Specifically targeting overlooked categories like festive wishes, this method allows users to customize spam detection based on contextual needs. By focusing on low-resource processing and dynamic adaptability, the framework effectively identifies clutter-inducing spam while (with 80% accuracy) remaining scalable for broader applications.

Semi-finished tobacco warehouses are critical nodes in cigarette manufacturing, where logistics capacity directly influences production efficiency. Under the framework of the Industrial Internet of Things (IIoT), this study develops a simulation-based optimization approach to enhance warehouse logistics capacity. An EMS (Electrical Monorail System) trolley scheduling strategy based on a “farthest-first” rule is proposed to improve track utilization and reduce congestion. One-way experiments and ANOVA confirm that both the number of lanes and trolleys significantly affect system performance. A full factorial design combined with multiple regression analysis produces a well-fitted (99.9%) agent model. Using the feasible direction method, the optimal configuration is identified as 11 lanes and 11 trolleys, with less than 2.5% deviation between fitted and simulated results. The proposed framework provides a quantitative and systematic method for logistics capacity design in IIoT-enabled smart warehouses, offering both theoretical insights and practical guidance for intelligent logistics applications.

This increasing dogmatism of the traditional wireless sensor network (WSN) systems poses severe problems in dynamic and resource-bound Internet of Things (IoT) systems. In this paper, a robust virtualization architecture is proposed to facilitate the deployment of virtual communities in IoT-enabled WSNs, enabling both node-level and network-level embedding. The core component of this framework is the ReVNE (Resilient Virtual Network Embedding) algorithm, which maximizes survivability, improves the rate at which virtual network requests are accepted, and reduces recovery delay after network failures have occurred. Large-scale simulations are used to test this solution and compare it to the existing algorithms, such as N-SVNE and RLE-SVNE. Results show that ReVNE is far more effective than current approaches in terms of throughput, latency, fault recovery, and routing overhead, thus providing a robust and scalable solution to virtualized IoT apps.

With the development of smart cities based on 6G technology, it greatly promotes the development of the tourism industry. Deep models based on time series forecasting have been widely applied in tourism economic prediction. However, these models cannot utilize heterogeneous data from multiple sources and are unable to model the long-range dependencies of time series. To resolve these issues, we construct an efficient variant spatiotemporal graph convolutional network for predicting the development of the tourism economy. Firstly, we collect heterogeneous data from different tourist attractions through 6G technology and construct tourist attraction graph data. Then we embed the Mamba module into the spatiotemporal graph convolutional network for long-range dependency modeling. Moreover, a gated recurrent unit (GRU) is introduced for temporal feature extraction. On our tourism economy forecasting dataset, experimental results show that our proposed model achieves the best predictive performance.

Fuzzy clustering algorithms have been widely used for complex data patterns and image segmentation using fuzzy partitioning. However, their performance degrades in the presence of noise and higher-dimensional spaces. To overcome these limitations, we propose a hybrid algorithm, DBKIFCM–PSO, which combines distance-based kernelized intuitionistic fuzzy C means clustering with particle swarm optimization. This hybrid approach addresses challenges in noisy datasets and higher-dimensional spaces, providing optimal solutions. We compare DBKIFCM–PSO with existing algorithms on various datasets and demonstrate its superior performance. By integrating DBKIFCM and PSO, our algorithm has shown improved performance on evaluation metrics like $��{\mathrm{VP}}_{\mathrm{PC}}�$, $��{\mathrm{VP}}_{\mathrm{PE}}�$, and $��{\mathrm{VP}}_{\mathrm{FS}}�$ over counterpart method by obtaining best values of 0.9595, 0.1036, and −61 120, respectively, on these metrics making it suitable for real-world applications.

This paper proposes AQOF-LM, an adaptive Quality of Experience (QoE) optimization framework empowered by large-scale pretrained models and multimodal data fusion through wireless communication platforms to promote population health. AQOF-LM consists of three key components: a QoE Forecasting Engine Based on Multimodal Semantics, a Context-Aware Exercise Classifier, and a QoE-Aware Transmission Controller. The forecasting engine utilizes video, sensor, and biometric signals to predict subjective QoE in real time. The exercise classifier identifies motion context to guide content encoding, while the transmission controller dynamically adjusts delivery strategies under network constraints using a utility-based policy with hysteresis switching. We implement and evaluate AQOF-LM on two public multimodal datasets under simulated wireless conditions. Experimental results demonstrate that AQOF-LM outperforms state-of-the-art baselines in prediction accuracy (MAE = 0.19), delivery stability (QoE stability = 0.92), and latency responsiveness (58 ms). Moreover, AQOF-LM maintains perceptual quality under bandwidth degradation, achieving up to 25% higher QoE in constrained scenarios. These results establish AQOF-LM as a scalable and context-sensitive framework for enhancing real-time health guidance services in community and low-resource environments.